A Knowledge based Navigation System with a Semantic Map Approach
Mizue Kayama and Toshio Okamoto
Hypermedia provides an effective activity environment, where users can acquire knowledge by exploring the hyperspace in their own way. Still, users often tend to "get lost'' in the hyperspace (Nielsen(1990)). To improve such undesirable effects on users, many researchers have been working to construct hypermedia systems which can identify the user's interests, preferences and needs, and give some appropriate advice to the students during the exploring activity process (De La Passardiere & al.(1992)), (Shneiderman (1989)) .
In this paper, we propose a framework based on a sub-symbolic approach for the support of exploratory activities/learning in a hyperspace. By using our framework, it is possible to express the semantic feature of whole hyperspace and the state of the exploratory activity in topological order. This approach is applied to generate the navigation information for exploratory activity. By using our approach, the space explored is reconstituted automatically by the semantic similarities of the nodes which constitute that space.
This shapes a kind of map. We call this map the semantic features map of hyper space, in short Hy-SOM. The Hy-SOM is applied to express the user model and navigation strategy for user. The exploratory history of the users is mapped on the Hy-SOM. Then the semantic relations between nodes are given on the map. The result shows the exploratory states of the user. This is interpreted as a user model. The exploratory tendencies of the user are inferred from the above user's exploratory model. Based on the deduction of the user's exploratory tendencies and expert knowledge embedded in the system, a navigation trying to improve the undesirable patterns is developed. The navigation strategy parameters are instantiated by using the Hy-SOM, the hypermedia node hierarchy and the semantic attributes of both nodes and links. The result of the instantiation is an ordered list of nodes recommended by the system for reference in the next step of the user's exploration. The system presents this ordered list to the user as navigation guidance information.
The entire process, starting from the logging of the exploratory search history, till the navigation guidance information generation, is based on the collaborative problem solving model. This model contains a collaborative memory based on a hierarchic structure, and implements the collaborative-problem-solving's result sharing mechanism.
2. Research Goal
The purpose of this study is to propose a framework to support exploratory learning based a-priori knowledge embedded in a semantic feature map of the hyperspace and a user model of the learner's exploratory activity. To support the exploratory learning, our navigation system realizes the following sub-goals :
Based on these sub-goals, our system gives the learner navigation guidance information. The functions are as follows. Firstly, we need a function that reflects the whole structure of the hyperspace composed of elements of teaching material with a hyper-structure. To bring this to fruition, we analyze the whole hyperspace from two points of view. One is the physical structure, related to the hierarchical structure of the node. The other is the semantic structure, i.e., the meanings between nodes in hyperspace. To express the semantic structure, a map, composed by the semantic features of the links between elements in hyperspace, called the Hy-SOM, is introduced. Secondly, a function to grasp the exploratory state of the learner who uses the teaching material, is needed. To realize this idea, we introduced two types of user models. On one hand, numerical indices of the state of the exploratory activity are used. Exploratory activity states are divided into five types : search, estimate, selection, reference and evaluation. The indices computes the accomplishment of each of these states. The second user model is a model of the exploratory activity as a partial graph of the Hy-SOM. Thirdly, a function to generate navigation information for the user, is necessary. By using two types of user model, as above, the system decides a priority for each of the navigation strategies, selects and orders the nodes which are appropriate for the current state of the learner and displays the navigation information. The navigation system has to adapt to the learner's current state of understanding of his/her exploratory activity. Especially, the process of selection of a navigation strategy and the selection and ordering process need to adapt to the learner's state.
3. Adaptive Hypermedia
As mentioned above, "getting lost'' is one of the undesirable aspects of the hypermedia (Nielsen(1995)). Additionally, there are some points which spoil the effects of exploratory activity. To improve these points, a lot of researches are reported (Calvi (1997)),(Conklin(1988)),(Da Silva (1998)). Adaptive Hypermedia (AH) is one of these kinds of researches(Brusilovsky (1996a)). The interpretation of the purpose of the user's past search activity is deducted from the structure of the hyperspace and the user's search history, adapting to the user. The methods for adaptation are of two types. One is the content-level adaptation . This method changes the contents of the node which the user will refer in next step. This type is also named adaptive presentation system. The other one is the link-level adaptation. This method changes the links in the current node. This is also named adaptive navigation support system. The ELM-ART(Brusilovsky (1996b)), which is an intelligent tutoring system, is one of the most famous adaptive presentation systems. Its adaptation was realized based on a specific domain knowledge with a symbolic approach. So, the generality of the adaptation mechanism and the flexibility of the functional extension of this system are not high. The Adaptive HyperMan(Mathe (1996)) is one of the most famous adaptive hypermedia systems using a method of the content-level adaptation. This system is an example of interactive adaptive hypermedia. This system bases its adaptability on the use of an Adaptive Relevance Network, which is made of a collection of personal data. This mechanism can store the users' interests. The exploring characteristics of each user are acquired through conversation with him/her. This system is useful for a situation in which a frequent interaction between the system and the user is needed.
The systems which have been reported by now can be classified into the following four groups, according to the usage. They are 1. Web-based ITS(CAI) systems, 2. Information retrieval systems, 3. On-line help/ information systems, 4. Exploratory activity states improvement systems. The goals of each supporting system are divided into two types. They are "the assurance of the effectiveness of the results of the activity (study)'' and "the support of the exploratory activity''. The object of the user modeling is different, respectively. In the case of the systems which are categorized in group 1(De Bra (1998)), the object is to understand the user's state. Some traditional methods for user modeling (like the overlay model) are used in most of the systems. In the case of group 2 (Kaplan (1993)), it is the user's preference for certain topics and his/her interests. This type of system is developed not only for individual activity, but also for group activity. In the case of group 3(Kashihara(1998)), the object of the user modeling is the purpose of retrieving/search. For group 4(Gaines(1995)),(Mukherjar (1995)), the object is the exploratory activity of user. Some systems express the user's exploratory activity by using some qualitative indices[(Kayama(1998)). Other systems offer activity environments which can control the degree of cognitive loads for users.
Our research belongs to group 4. Our navigation guidance is based on the expertise about supporting exploratory activity. This expertise is navigation knowledge. For example, in the field of education, Educators use this kind of knowledge not only to design their instructional plan, but also the learner's achievement according to the learning goal. The educator defines the semantic attributes of each content in the educational hypermedia based on his/her instructional plan. Based on these data, the system creates the semantic feature map of the hyperspace. We have developed an adaptive navigation system for exploratory activity by using a sub-symbolic approach (Kayama (2000)). This system generates an advice message for a user who is in impasse, and also provides a list of recommended nodes which the system proposes to the user to refer as the next step of exploratory activity. The system shows advice to the user to give the student a chance to escape from undesirable exploratory states. To increase support effect for the learner, the system's navigation function is being improved, to attach semantic knowledge to whole elements in hyperspace. In this study, we use the relations between the concepts, which each node in the hyperspace represents, as the preliminary knowledge, for the process of navigation information generation. By using this knowledge, the system develops a model of the user's exploring activity on a semantic feature map of the hyperspace. The system generates navigation information which ensures that the exploring activity becomes reasonable and continual.
4. Semantic Feature Map of the Hyperspace : Hy-SOM
In our approach, Hy-SOM is used to express the characteristics of semantic relations between nodes in the hyperspace. Our technique bases on the original Kohonen self organizing feature map (SOM) (Kohonen(1990)).
The SOM is one of the most famous neural networks with unsupervised learning, used for classification in general. It works as a function which maps the topological characters of the training data on the two-dimensional (or three-dimensional) space (Hassoum (1995)), (Kangan(1990)),(Yin (1995)). This function is applied to create the semantic feature map of the hyperspace from the semantic relations of all nodes. As a result, the map of the nodes in the hyperspace is obtained, with the semantic similarities of each node maintained. The trained map is used to develop a learner's exploratory activity model, and to generate the navigation information for the learner. The exploratory history of the user is mapped on Hy-SOM. Then, the semantic relations between nodes are given on the map. The result shows the exploratory state of the user, interpreted as a user model.
By using Hy-SOM, it is possible to express the similarity of each topic and each node. Moreover, there is a possibility that the tendency of the exploring activity of a learner can be found by interpreting the semantic relations between the referred nodes.
Two extensions of the original SOM are proposed. Firstly, for improving the accuracy of the ability to classify, a new learning function is proposed. With this extension, the features of the input pattern reflect the structure of the output layer more exactly. Concretely, improvements based on the theory of probability are made on the composition of the input pattern and the initial phase of the learning (training) process. Figure 1 ((a) and (b)) shows the effect of our method. Categories for distinguishing each trained pattern (node) are created in the trained network. The configuration of the nodes is based on their semantic topological similarity. Secondly, for improving the robustness of the ability to classify, a method of reconstructing all weights in the trained network is proposed. In this way, some regions are defined on the map. Each region shows distinct topics (semantic cluster). Concretely, in order to represent visually the appearance probability of the value of each element in the input pattern (forming a vector), the trained weights are reconstructed into binary values. The topic configuration, similar to the node configuration, is based on semantic topological similarity. As a result of this structure, semantic similarities between topics, of which the course designers or instructors may not be aware, can appear clearly expressed on the map automatically. The lower part of Figure 1 ((b-1) and (b-2)) shows the effect of this extension.
By using our approach, the space explored is reorganized automatically according to the semantic similarities of the nodes which constitute that space. The resulting shape is a kind of map. We call this map the semantic feature map of the hyper-space, in short Hy-SOM. The Hy-SOM is applied to express the user model and navigation strategy for user. The exploratory history of the user is mapped on the Hy-SOM. Then, the semantic relations between nodes are given on the map. The result shows the exploratory state of the user. This is interpreted as a user model.
5. Hy-SOM based Navigation for Exploratory Activities
The exploratory history is used to express the states of the exploratory activity of the user. The history is expressed as a set of attributes of the referred nodes and the selected links. To identify the referred nodes, the system records sets of the type of contents and particular number of the node. After a learner refers a node, the system asks the learner, whether this node is valuable for his/her exploratory goal or not. Then, the answer from the learner is recorded as one item of exploratory history data. This data is used as an index which shows whether the selection of a node is suitable for the exploratory intention of the user. To recognize the selected link, the system records two types of data. They are the particular number of the link and the type of relation between starting node and terminal mode.
5.1 A Learner, demanding Navigation Guidance
Based on the result of our pre-research, the situations in which a learner demands the navigation guidance information are put together as follows.
As navigation guidance, the learner requests two kinds of information. One is the topic, the other one is the node, which the learner should explore next.
5.2 User Modeling on the Hy-SOM
The learner's exploring activity model is expressed by using Hy-SOM. The exploratory history is arranged to map on Hy-SOM. Some unimportant actions, for example "BACK'' and "FORWARD'' and "GO to History'', are excluded from the learner's history. Then, the type of each link (type of semantic relation and node hierarchical structure) is appended to the history. As a result of this arrangement, the history is mapped on Hy-SOM.
The user model assembles information of two kinds. The first type of information is extracted from the exploratory history of the Hy-SOM. The second type of information is gathered from the interpretation of the exploratory history on the semantic network built from the attributes of the nodes and links. From these two types of information, our navigation system creates the user's exploratory activity model. The user exploratory model has three states: reference state, exploratory state and cognition state. The exploratory tendencies of the user are inferred from the above user's exploratory model. Based on the deduction about the user's exploratory tendencies and on the expert knowledge embedded in the navigation system, navigation guidance trying to correct the undesirable patterns is developed. The navigation strategy parameters are instantiated by using the Hy-SOM, the hypermedia node hierarchy and the semantic attributes of both nodes and links.
The result of the instantiation is an ordered list of nodes recommended by the system for reference in the next step of the user's exploration. The system presents this ordered list to the user as navigation guidance information.
5.3 Navigation Knowledge
The navigation knowledge which the educator(s) possesses is defined as follows. It is expertise concerning the characteristics, to which an expert pays attention to, when the applied strategy to support the learner is decided. Here we make one assumption, to define the navigation knowledge. When an expert selects the appropriate navigation strategy for the learner, he/she notices a particular value or a set of values. To extract the navigation knowledge from educator(s), we have to make investigations for educator(s) with the experience to develop their class using educational hypermedia. The instructional strategy for the exploratory learning support is arranged, based on the result of that investigation. These strategies are the navigation strategies to generate guidance information. Then, we collect the values (or set of values) of the exploratory state indices which represent the characteristics that determine the application of one or another strategy.
5.4 Navigation Strategies
The strategies to generate navigation information given to the learner are arranged in consideration of the relation between the navigation knowledge of educators, the exploratory history of the learner, the current node or the current topic and some other information from the structural model of hyperspace. The method of node selection and giving priorities for each strategy is as follows.
The systematization of the exploration(Se) : Systematic exploration is supported. On the Hy-SOM, the node which is located around the current node is presented. This is the strategy applied in order to stimulate the systematic understanding of the contents of the current node.
The diversification of the exploration (De) : Diversified exploration is supported. The node with a semantic relation to the current node is presented. This strategy is applied in order to stimulate the diversified understanding of the contents of the current node.
Preventing exploratory shortage (PeS): The exploration is supported for the achievement of the exploratory goal of the learner. In the referred topics, non-referred nodes with high importance degree are presented. This strategy is applied in order to make sure that nodes with important contents are referred. When a strategy is applied, the priority of each node is decided in proportion to the distance to the current node on the Hy-SOM.
The extension of the exploratory viewpoint (EeV): This strategy supports the extension of a learner's viewpoint of exploration. On the Hy-SOM, a topic with many semantic relations to the current topic is presented. This strategy is applied in order to stimulate the diversified understanding of the current topic.
The focusing of the exploratory viewpoint (FeV): This strategy supports the focusing of the learner's viewpoint of exploration. Within the topics which the referred node belongs to, a topic with the highest average of the degree of importance of all un-referred nodes are suggested. This strategy aims at the establishment of a knowledge on the specific topic.
The conversion of the exploratory viewpoint (CeV): This strategy supports conversion of the viewpoint of exploration. On Hy-SOM, a topic which is located around the current topic is presented. This is a strategy that leeds to being able to discern the important topics, and to appropriately estimate the value of the concerned topic. When a strategy is applied, the priority of each node is decided in proportion to the distance to the current topic on the Hy-SOM.
6. System Architecture and the Interfaces
The entire process, starting from the logging of the exploratory search history, till the navigation guidance information generation, is based on the collaborative problem solving model. This model contains a collaborative memory (Quillian(1968)) based on a hierarchic structure, and implements the collaborative-problem-solving's result sharing mechanism.
In Figure 2, the architecture of our navigation system is shown. This system consists of two sub-systems. One is the Hy-SOM computing system. The users of this system are the designers of the educational hypermedia. They have to embed the semantic attributes of whole nodes and links in the hypermedia. Our system also offers an attributes editor. The other sub-system is the navigation information generation system. The interfaces of this system are shown in Figure 3. Three types of navigation guidance information (recommendation of the appropriate nodes to explore in the next step, the whole map of the educational hypermedia and the learner's exploratory history) are given.
In this paper, we have presented an overview of our semantic feature map-based system for exploratory activity support. Our approach is an atempt to increase the validity of navigation by arranging the structure of the hyperspace from the viewpoint of the hierarchical structure of nodes and of the semantic relation of contents. By creating a knowledge based semantic feature map as a hyperspace model of the semantic structure, it becomes possible for the system to express not only the similarity between nodes, but also the similarity between topics. Educator(s) can not be aware of these features beforehand. The topological concept of the node (or topic) is introduced in the process of generating navigation guidance information. In this way, we expect that the validity and suitability of the navigation information are improved. The training patterns for creating the Hy-SOM have influence on the accuracy of the semantic feature map. The training pattern is a vector which is made of semantic features of a node in the hyperspace. Therefore, the method to develop this vector must be analyzed by considering the processes which create the navigation guidance information for the user. This aspect needs future detailed examination, and constitutes a part of our future research.
We have focused especially on the two extensions we proposed for our Hy-SOM, for improving both classification accuracy and robustness. Moreover, we have shown the information types that we use for our user model. Finally, the system architecture was described.